Use case · Quality

Quality prediction & virtual metrology

Predict product quality from live process data so problems are caught before parts are measured, tested or shipped.

Maturity
Proven, scaling
Complexity
High (our assessment)
Department
Quality
Technology
Machine learning, Industrial IoT & edge
Solves
Scrap & rework, Defects reaching customers, Throughput bottlenecks, Long lead times

What it is

Quality prediction, sometimes called virtual metrology or soft sensing, uses machine learning to estimate a quality result from the process data recorded while the part or batch was made. Instead of waiting for a lab test, a CMM measurement or end-of-line test, the model predicts the outcome from temperatures, pressures, speeds, cycle profiles and raw material properties, often within seconds of the process step finishing.

It matters because physical measurement is slow, expensive or destructive, so most plants only sample. Defects made between samples can travel a long way down the line before anyone knows. A reliable prediction for every unit lets teams stop a drifting process early, measure smarter by focusing physical checks where the model is uncertain, and hold suspect material before more value is added.

How it works

  1. Choose a measurable target

    Pick a quality characteristic that is already measured and matters, such as a critical dimension, tensile strength, viscosity, moisture or a final test result. That historical measurement becomes the training label.

  2. Align process data to results

    Each measurement is matched to the exact process conditions that produced it, using serial numbers, batch IDs or timestamps adjusted for material transit time. This traceability work is often the largest part of the project.

  3. Build and validate the model

    Models learn the relationship between process signals and the quality outcome, then are tested on later production periods they did not see during training, to confirm they hold up across shifts, lots and seasons.

  4. Run alongside measurement

    The model runs live in shadow mode next to the existing sampling plan. Engineers compare predictions with actual results before anyone acts on predictions alone.

  5. Act on predictions

    Once trusted, predictions can trigger holds, extra inspection, or operator adjustments, and can justify reducing physical sampling where the model is consistently confident.

Benefits

Every unit gets a quality estimate

Sampling plans check a fraction of output; a prediction covers each part or batch, so a drift between samples is visible when it starts rather than at the next scheduled check.

Faster feedback to operators

Operators see a likely quality result within the cycle rather than hours later from the lab, which shortens the loop between a process change and its effect.

Less scrap from late detection

Suspect material is held at the step where the problem occurred instead of being discovered after downstream processing, packaging or shipment.

Smarter use of the lab

Physical tests and measurements can be targeted to units where the model is uncertain or the stakes are highest, rather than spread evenly.

Data & prerequisites

Data you will need

  • Historical quality measurements from the lab, CMM, gauges or end-of-line test
  • Time-series process data from the historian, PLCs and machine controllers
  • Serial, batch or lot genealogy linking process data to quality results
  • Raw material certificates and incoming inspection data
  • Equipment state, tool changes, maintenance events and recipe or setpoint changes

Before you start

  • Traceability good enough to tie a quality result to the conditions that produced it
  • Enough historical measurements that include both good and out-of-spec outcomes
  • Process engineers who can judge whether the model's drivers make physical sense
  • A defined procedure for what happens when a prediction says a unit is suspect

Software for quality prediction & virtual metrology

Companies whose published products address this use case. Listings are editorial, not paid. Filter the directory.

Eigen Innovations

Vision inspection & quality AI · Fredericton, Canada

Eigen Innovations combines thermal and optical imaging with AI to inspect parts and monitor heat-driven processes such as plastic welding, molding, adhesive application and furnace operation. Manufacturers use it for...

Visual quality inspectionQuality prediction & virtual metrologyPredictive maintenance
Founded 2012

Fero Labs

Process analytics & optimization · New York, NY

Fero Labs makes machine learning software that process engineers use to find root causes of production problems, simulate setpoint changes and run optimization on live production. It focuses on heavy process industries...

Root-cause analysis for quality & yieldProcess & yield optimizationQuality prediction & virtual metrology
Founded 2015

Real-world examples

Each example is labeled by who reported the result. How we grade evidence.

Supporting research

Artificial Intelligence with Open and Scaled Data Sharing in Semiconductor Manufacturing - Workshop Report

National Institute of Standards and Technology (NIST AMS 100-72) · Report · 2025

Report of an NSF-sponsored, NIST-supported workshop in which a semiconductor industry coalition led by Seagate and UCLA tested pooling and consistently processing data from multiple plasma etch machines across factories for an ML metrology application. It benchmarks the economic value of 'AI-ready' data and sets out data-first business principles for collaborative data ecosystems.

Government

2025 State of Smart Manufacturing Report

Rockwell Automation · Survey · 2025

Rockwell Automation's 10th annual global survey, produced with Sapio Research, of 1,560 manufacturing decision-makers in 17 countries on smart manufacturing investment, AI use, workforce, cybersecurity and sustainability. As a vendor report, it frames AI as a response to labor shortages, quality pressure and external risk.

Vendor

Artificial Intelligence in Advanced Manufacturing: Current Status and Future Outlook

ASME, Journal of Manufacturing Science and Engineering · Peer-reviewed paper · 2020

Arinez, Chang, Gao, Xu and Zhang review applications of AI and machine learning to representative manufacturing problems, organized by the hierarchy used in plants from the overall system down to material process streams. Topics range from throughput and quality to human-robot collaboration, process monitoring, diagnosis and prognosis, and materials process modeling and control.

Academic (peer-reviewed)

A review of artificial intelligence applications in manufacturing operations

Journal of Advanced Manufacturing and Processing (AIChE/Wiley); record via U.S. DOE OSTI · Peer-reviewed paper · 2023

Plathottam, Rzonca, Lakhnori and Iloeje review current developments, challenges and future directions for AI and machine learning in manufacturing operations. The paper aims to help practitioners choose suitable AI/ML technologies and to identify research areas with high potential returns.

Academic (peer-reviewed)

Machine Learning in Manufacturing towards Industry 4.0: From 'For Now' to 'Four-Know'

Applied Sciences (MDPI); repository copy at KITopen (Karlsruhe Institute of Technology) · Peer-reviewed paper · 2023

Chen and co-authors propose a roadmap for developing ML applications in manufacturing, from ideation to deployment, based on the published literature. They frame ML tasks along 'Four-Know' (what, why, when, how) and 'Four-Level' (product, process, machine, system) dimensions and summarize supervised, semi-supervised, unsupervised and reinforcement learning methods with typical applications.

Academic (peer-reviewed)

Common pitfalls

  • Weak genealogy, so the model learns from mismatched process and quality records
  • Training on a period with almost no defects, leaving the model unable to recognize bad outcomes
  • Cutting physical sampling too early, before predictions have been validated over enough operating conditions
  • Models that silently degrade after a recipe, material or equipment change because nobody monitors prediction error

Questions to ask vendors

  1. How do you handle the time lag between a process step and the quality measurement it affects?
  2. How do you show which process variables drive a prediction, so our engineers can check it makes physical sense?
  3. How do you detect when model accuracy is degrading, and what triggers retraining?
  4. What validation evidence will you provide for regulated or customer-audited processes?
  5. Can the model run at the line within our cycle time, and does it write results back to our MES or quality system?

What to measure

  • Scrap and rework cost at and downstream of the predicted step
  • Prediction error against physical measurement
  • Time from defect creation to detection
  • Physical inspection or lab tests per unit
  • Customer quality escapes

Funding that can help

Frequently asked questions

What is virtual metrology in manufacturing?

Virtual metrology predicts a measurement, such as a film thickness or critical dimension, from process and equipment data instead of physically measuring every unit. It originated in semiconductor fabs where metrology is slow and costly, and the same approach is now used in chemicals, plastics, metals and food. Physical measurement continues on a sample basis to validate and recalibrate the model.

How accurate does a quality prediction model need to be?

It needs to be accurate enough for the decision it supports. A model that only flags units for extra inspection can tolerate more error than one used to release product without testing. Teams usually start by using predictions to prioritize inspection and only reduce sampling after validating performance over a wide range of conditions.

What is the hardest part of building a quality prediction model?

Usually it is data alignment rather than the model itself. Every quality result has to be matched to the exact process conditions that created it, which requires good serial or batch genealogy and correct handling of transit times between steps. Plants with weak traceability often need to fix that before prediction is reliable.

Related use cases

Quality

Root-cause analysis for quality & yield

Machine learning sifts process, material and equipment data to narrow down which factors are actually driving defects and yield loss.

Machine learningGenerative AI & LLMs
Proven, scaling8 software companies

Operations & production

Process & yield optimization

Models recommend or automatically adjust process setpoints to keep quality on target while improving yield, throughput and energy use.

Machine learningOptimization & planningIndustrial IoT & edge
Proven, scaling7 software companies

Operations & production

Process anomaly detection

Models learn normal process behavior across many sensors and flag unusual patterns before they become quality, safety or equipment problems.

Machine learningIndustrial IoT & edge
Proven, scaling7 software companies

Quality

Visual quality inspection

Cameras and computer vision models inspect parts on the line to catch surface, assembly and dimensional defects consistently.

Computer visionMachine learning
Proven, scaling8 software companies

Put AI to work

Find the AI opportunities hiding in your plant.

Genvo, our implementation partner, works with manufacturers to pick the right first project, connect the systems you already run and put AI agents and tools into daily use, with people approving what goes out.

Assessments and consultations are provided by Genvo. How our partnership works.